EMBARCADERO LABS · AGENTIC PHYSICAL AI
What we're about
Embarcadero Labs builds agentic physical AI: AI that guides work in the physical world in real time, learning from the people who do it best.
We capture expert work from the worker's own point of view, as it happens. We check every step against something a meter can confirm, and we use the runs that pass to guide whoever comes next. The trades are losing experienced people faster than they can train new ones, and almost nothing records what those people know. That's where we start: field work today, and the data that teaches agentic physical AI tomorrow.
Five ideas behind it
1. The know-how isn't online.
Language models learned from the internet, but the physical world isn't on it. A model can recite a cooling tower's manual, but it doesn't know site-specific details, and it can't see the local context: which unit keeps failing, what was tried last time, what's in front of the worker right now. That knowledge lives with the people who do the work, and paperwork doesn't scale it. Video from the worker's own point of view, recorded on real jobs, does. It's also the format robots learn from.
2. Proven runs are learnable.
A proven run is a real job, recorded from the worker's point of view, where every step passed a check a meter can confirm and the fix lasted: the unit didn't come back. In AI, passing a check like that is a verifiable reward, the recipe behind today's reasoning models. So the run to follow is the one that proved itself, not the most senior person's. The corrections and exceptions along the way are the most valuable data of all.
3. Agentic physical AI is a loop.
Physical AI can't answer once and stop. It sees, decides, acts, and looks again: vision, model, action, world. Today a person closes that loop. AR glasses are the eyes, Ten4 and the expert's proven run are the model, and the worker's hands act. Some of the loop can't wait for the cloud. When wind starts to turn a locked-out fan, the warning has to come from the glasses themselves.
4. Fast hands, slow plans, and the part in between.
Kahneman split thinking into a fast, automatic System 1 and a slow, deliberate System 2. AI is splitting the same way. World models and robot policies keep getting better at System 1: short, small-scoped actions over a second or two, like a hand finding a belt. Language models, given enough context, are good at System 2: reading the manual and planning the job. What's missing is the connective tissue between them: knowing where the job stands right now, what comes next, and when the plan has to change because the fan just moved. That's spatiotemporal reasoning, reasoning about space and time together, and it's what we mean by agentic physical AI. Our part is the ground truth for that connective tissue: step-by-step records of where real jobs stood, labeled by whether they worked.
5. Start with field work. Scale to agentic physical AI.
AI is climbing a ladder: from models that talk, to models that see, to world models that predict how physical things behave, to agentic physical AI that acts in the world. Each rung needs data the rung below never had. That data comes from the field. Swapping a belt is easy for a person and still hard for a robot, so we help people first. The same proven runs that guide a crew today are the kind of data that could teach machines later.
The product: Ten4
Ten4 is an iOS app + cloud platform for contractors, an AR-enabled AI assistant for frontline trades workers. ten4.works ↗
Press and recognition
- Inc. Meet the former NVIDIA insider seeding AI into small businesses ↗
- Meta A white paper on AR for frontline workers, published with Meta ↗
info@embarcaderolabs.com · Austin, TX & New York, NY